Why Shorter Prompts Work Better: Lessons from OpenAI’s GPT‑5.6 Guide

OpenAI recently released a GPT‑5.6 prompt guide that flips the long‑standing "more is better" mindset in prompt engineering. The guide’s top recommendation—"shorten your prompts"—is backed by internal coding‑agent tests where removing repeated instructions and excess examples raised scores by roughly 10‑15% and cut token consumption by 41‑66%, slashing costs by 30‑60%.

Why longer prompts hurt

Older models required repeated phrasing to be understood, leading developers to pile on role definitions, examples, and defensive commands. Newer models have stronger intent comprehension, so redundant lines become noise: repeating "don’t modify tests" makes the model over‑cautious, and duplicated commands distract rather than reinforce compliance.

How to trim prompts safely

Start with a working prompt and a toolset. Delete one group of instructions or examples at a time, then rerun the same evaluation. If the score does not drop, keep the deletion and continue. The core principle is to state each instruction only once and keep tool descriptions brief and precise.

Do not delete required examples (e.g., output format specifications) or proven style guidance that demonstrably improves scores. Excessive content in long conversations also inflates token counts each turn, degrading performance.

Skeleton for concise prompts

The guide provides an eight‑section checklist (role, personality, goal, success criteria, constraints, tools, output, stop rules). Each item should be one or two sentences that directly affect model behavior. For a Go code‑review assistant, the prompt looks like:

角色：你是我们组 Go 后端仓库的 code review 助手。
性格：直接指出问题，不用客套。
目标：审查这次 PR 的改动，给出可执行的修改意见。
成功标准：每条意见指向具体文件和行号，说清为什么是问题。
约束：只看这次改动涉及的文件，别顺手重构无关代码。
工具：可以跑测试和 lint，不许改代码。
输出：按严重程度列出问题，最多十条。
停止规则：信息不够下判断时，列出你缺什么，别猜。

Use this skeleton as a ruler: any existing prompt line that does not map to one of the eight sections can be removed.

Granting the model autonomy

The guide recommends three tiers of actions: (1) read‑only tasks can be performed freely; (2) reversible local changes (e.g., modify code, run tests) are allowed without asking; (3) external, destructive, or costly actions require explicit confirmation. Contradictory instructions (e.g., "never modify tests" vs. "fix failing tests") confuse the model and degrade performance.

Instead of hard‑coded "ALWAYS" or "NEVER" rules, use clear, essential constraints (e.g., "NEVER commit secrets"). Ambiguous judgments should be expressed as a decision standard that prompts the model to ask for clarification when needed.

Ensuring results are verified

Generating an answer does not mean the task is complete. The guide advises attaching verification steps: provide tools for testing, type‑checking, or building, and specify which validations are critical. For front‑end work, also manually inspect rendered pages for layout issues.

Verification criteria must be part of the acceptance standard—state how to prove completion, otherwise the model’s claim of "done" is untrusted.

Writing concise answers

Because GPT‑5.6 already produces terse outputs, adding prompts like "be brief" can over‑compress important information. Instead, tell the model what must remain in a short answer: the conclusion, supporting evidence, key cautions, and next steps. Remove only secondary details and repetitions.

Similarly, avoid vague tone modifiers like "be friendly"; describe the desired behavior directly (e.g., "give the answer, then only add reassurance if truly needed").

Final thoughts

The shift in prompt engineering is from lengthy, rule‑heavy scripts to contract‑style prompts that clearly define goals, boundaries, and acceptance criteria while letting the model decide the execution path. As models become more capable, each unnecessary word costs more tokens, so the ability to prune prompts becomes a key skill.